Researchers poison their own data when stolen by an AI to ruin results (www.techradar.com)

Researchers from China and Singapore have developed AURA (Active Utility Reduction via Adulteration), a novel defense mechanism designed to protect proprietary knowledge graphs used in Generative AI systems like Microsoft's GraphRAG. By intentionally poisoning these knowledge graphs, AURA ensures that any stolen data yields inaccurate outputs and hallucinations. To retrieve correct answers from the system, users must possess a secret key, making unauthorized access nearly pointless. The researchers demonstrated AURA's effectiveness, achieving approximately 94% success in degrading the utility of any stolen knowledge graphs. This development holds significant implications for the AI/ML community, particularly as knowledge graphs become increasingly integral to generating accurate responses in AI applications. AURA offers a practical approach to safeguarding intellectual property against cybercriminals and other malicious actors who might exploit these valuable datasets. As AI models continue to rely on real-time, proprietary information for enhanced decision-making, AURA could be crucial in maintaining data integrity and ensuring that sensitive knowledge is better protected against theft and misuse.